isValidUtf8 is broken about bytestring

[Bodigrim]

Yeah, that's my conclusion as well (I've been looking at aarch version). Seems wrapping lookahead manipulation in if(big_strides) should be enough.

GHC team has asked for any version bumps to be included in GHC 9.4.8 to be ready by Friday: https://mail.haskell.org/pipermail/ghc-devs/2023-October/021420.html. We probably want to fix this and backport to bytestring-0.11.

@clyring if you get time to put up a PR, fill free to merge and release without my approval, I'll be AFK until next week.

[sol]

For the minimal case isValidUtf8 (drop 1 "\128\0")

• both is_valid_utf8_avx2 and is_valid_utf8_ssse3 are broken
• while is_valid_utf8_sse2 and is_valid_utf8_fallback seem to work

[clyring]

I see the problem. Here is the relevant bit of the avx2 code:

  // 'Roll back' our pointer a little to prepare for a slow search of the rest.
  uint32_t tokens_blob = _mm256_extract_epi32(prev_input, 7);
  int8_t const *tokens = (int8_t const *)&tokens_blob;
  ptrdiff_t lookahead = 0;
  if (tokens[3] > (int8_t)0xBF) {
    lookahead = 1;
  } else if (tokens[2] > (int8_t)0xBF) {
    lookahead = 2;
  } else if (tokens[1] > (int8_t)0xBF) {
    lookahead = 3;
  }
  uint8_t const *const small_ptr = ptr - lookahead;
  size_t const small_len = remaining + lookahead;
  return is_valid_utf8_fallback(small_ptr, small_len);

When the input is too small for a 128-byte big stride, we reach this code with prev_input containing all zeros.

This lookahead fiddling is meant to ensure we do not call the fallback after just part of a multi-byte code point. But for some reason it looks for any non-continuation-byte in prev_input rather than specifically for valid first bytes of a multi-byte code point. And zero is not a continuation byte, so we end up checking if [start-1,end) is valid rather than [start,end) as intended.

I'll put up a PR soon.

[clyring]

It seems I lack the authority to circumvent the approval requirement before merging to master.

Well, I pushed to bytestring-0.11 and cut a release anyway.